Before spending billions on more scale, we should ask whether the bottleneck is the model itself, or the way we’re asking it to work. There may be more gains in changing the way engineering happens rather than expanding the frontier. https://t.co/HLC4iBWymx
@kirtandopamine Nah, I'm post-markdown at this point.
Engineering sketches of system internals in a compilable form are 10x better. You can see what's going on and police the architecture quite easily without suffocating under a pile of markdown slop.
@danpdc The debates on X are reductive beyond belief.
"Anyone still reading code in 2026 is a luddite"
"If you arent vibe coding everything you're behind"
"AI is for crappy devs that only generate slop"
"There will be no more jobs in 12 months"
Dunno know how much longer I can take it
My productivity is easily 10x, even 100x in some cases.
I ran a devtools company for 10 years. Now I design languages, type systems, write compilers. Perhaps I am a second-rate programmer to you.
I had to make significant changes to my workflow to get my gains. I now design the system internals, the class shapes, types, etc, and hand the implementation off to the model.
@ClaudeBindels A lot of the good ones are just private get-togethers. The ones you find on Luma are what I call the "leaves". They're a good starting point, usually they're 49 generic people and 1 interesting person who intros you to better circles, aka the "branches".
@giadotai Well to add one data point to this... I've basically dropped Codex now for DeepSeek (i do compiler dev work)
And I use Flash, not Pro, just prove the point that Frontier tokens are silly and almost nobody needs a 900 quadrillion parameter model.
@ivanburazin Many, many UIs are going to go away. This is a good thing.
On the chat log side tho, we still haven't really seen hybrids that are chat logs with that have rich UIs integrated within beyond embeds. Once that gets solved I suspect it will become a major UI paradigm.
@azamsharp I do research professionally on AI coding strategies. You can get vastly better results from AI if you design a sketch of the system internals in an architectural format, and then hand that off for implementation to the model.
@xoaanya I'm daily driving DeepSeek Flash 4.1. If you need better perf you can move up to Pro but for 99% of tasks Flash is fine.
And my workload is doing frontier work on compilers and coding agents, and it's even handling that fine.
@stevesi Yes, but the “no, this time really is different” argument has some merit here. After all, there’s considerable money to be made by advancing patently absurd (and perhaps irresponsible) claims about the current state of AI cybersecurity.